Simultaneous Perturbation as a Spectral Filter

Abstract

Parameterized quantum circuits require repeated measurements to estimate objective gradients, making the number of circuit evaluations a central resource. Simultaneous perturbation stochastic approximation (SPSA) is widely used because it obtains a full gradient estimate from only two cost evaluations, independently of the number of parameters. We analyze the random perturbation for the structured Fourier landscapes generated by quantum circuits. For a finite update width, we derive the exact mode-by-mode response of the mean Rademacher estimator and show that it selectively suppresses modes involving many parameters. For general generators, the filtered mean field need not be conservative. We derive its sampling variance, compare it with Gaussian random-direction smoothing, and formulate the resulting stochastic dynamics through a Fokker--Planck equation. The non-conservative field generates acceleration to the steady state and the global minimum of the cost function by gradually decreasing the parameters in SPSA.